AMD Radeon RX 9070 XT is a 304 W TBP Card, Regular RX 9070 Comes with 220 W Configuration

[Daven]

Yeah that’s a seriously big power gap that stood out to me too

The 9070 (56 CUs) is 220W. The 9070XT (64 CUs) is 300W. Difference 80W.

The 4070S (56 CUDA cores) is 220W. The 4070TiS (66 CUDA cores) is 285W. Difference 65W.

I don't think the gap is different enough compared to other examples to warrant any concern.

 

[alwayssts]

This is actually somewhat concerning to me. This probably means that the card starts hitting diminishing returns at the 200-250W range and AMD is only ramming +300W through the XT model to hit a certain performance milestone.

It's still probably 350W with max power limit. I won't explain how power/voltage works, because I know people think it's boring, but the chip *could* even use a lot more (for high clocks).
Hope those coolers are something special (if they plan on ramping voltage/clocks >375w), though!

Again, as I've said too many times, it makes sense to have an *up to ~375w* xt and a >375W xtx for greater clocks (and potentially use 24gbps memory, if there isn't enough cache). This truly is the big question.

Either AMD wants people to be surprised, or this chip needs to be cheap bc it would be very unexciting.

 

[AusWolf]

304 seems weirdly specific, like, they could have just said 300 or 305 and we'd have read the delta as margin of error or something, not sure why they felt the need to give out such a precise number...

They started it with the 7800 XT which was a 263 W card, god knows why.

besides that, I also saw the news about XFX's 9070XT magnetic air and it's looking real fine to me ! can't wait to snag a white one for my build !

I love the idea of interchangeable fans in the Magnetic Air series, although my attention is on the PowerColor Reaper this time, which seems to be the only truly dual-slot 9070 XT.

 

[alwayssts]

Yeah that’s a seriously big power gap that stood out to me too

Artificial product segmentation says "hola". 9070 is clearly meant to beat/compete with 5070 and *that's it*. Pssst....that's how you know it's going to be cheaper than you may think.

Everything else is for higher-end cards. At some point people may understand what I've been saying for months.

This is because you could *potentially* use all the bandwidth from 20gbps ram, that both 9070/9070xt use, with just a 7168sp model at high clocks...therefore making the 9070xt model redundant? Make sense?

I don't like it. Let the chip have it's full potential. If they need to cut it down, cut it down, but then sell it cheaper! I think imposing artificial limitations (to this potential extreme) is fucking lame and if it's true, sucks.

They started it with the 7800 XT which was a 263 W card, god knows why.


I love the idea of interchangeable fans in the Magnetic Air series, although my attention is on the PowerColor Reaper this time, which seems to be the only truly dual-slot 9070 XT.

7800xt was literally limited by power (to ~45TF); it overclocked literally 20%, that's why it used it. If that card could have drawn more power (and it's RT didn't suck), there would be no this chip.
Again, it's alllllll artificial product segmentation by limiting voltages to the core/memory (obviously nothing could save Navi3 RT, though).

 

[AusWolf]

Artificial product segmentation says "hola". 9070 is clearly meant to beat 5070 and *that's it*. Everything else is for higher-end cards. At some point people may understand what I've been saying for months.

I don't like it. Let the chip have it's full potential. If they need to cut it down, cut it down, but then sell it cheaper! I think imposing artificial limitations (to this potential extreme) is fucking lame and if it's true, sucks.

I'm fine with it. I don't need or want a card for every single FPS number between 80 and 120 like Nvidia seems to think these days.

7800xt was literally limited by power (to ~45TF); it overclocked literally 20%, that's why it used it. If that card could have drawn more power (and it's RT didn't suck), there would be no this chip.
Again, it's alllllll artificial product segmentation by limiting voltages to the core/memory (obviously nothing could save Navi3 RT, though).

While the 7800 XT could draw a lot more power, I don't think there was a point. I had one for a brief amount of time, and while you could easily just max out its power slider at +20%, it only increased GPU clocks by about 100 MHz which didn't result in any perceptible performance difference (a couple more 3DMark points mean nothing in my books), but increased heat and noise by a lot.

 

[alwayssts]

I'm fine with it. I don't need or want a card for every single FPS number between 80 and 120 like Nvidia seems to think these days.


While the 7800 XT could draw a lot more power, I don't think there was a point. I had one for a brief amount of time, and while you could easily just max out its power slider at +20%, it only increased GPU clocks by about 100 MHz which didn't result in any perceptible performance difference (a couple more 3DMark points mean nothing in my books), but increased heat and noise by a lot.

The first part is LITERALLY WHAT THEY'RE PROBABLY DOING. Cannot confirm. We will see, but thinking it's likely the case and why sad. That's the point. That said, it could make lower product(s) cheaper...maybe.

That's what guys need to understand. Either higher-end product will exist (at some point) and/or it will be cheap, if not both. If they're trying to wring out every last cent, well then that just sucks for us.
That's not usually why you build a stack like this. It is (conceivably) the 4850/4870/4890 stack. 4850 was cheap as hell...like the equivalent of $400. THAT'S WHAT I'M SAYING. That's why low-power/clocks.

If your 7800xt only clocked to 2530mhz, you were likely doing it wrong and/or had the worse chip I have ever seen (remember the 12% in that graph is on TOP of a factory OC). That's all I can say about that.

 

[AusWolf]

If your 7800xt only clocked to 2530mhz, you were likely doing it wrong and/or had the worse chip I have ever seen. That's all I can say about that.

I don't know. I'm not big on overclocking anyway. Playing with the clocks only made them a bit more unstable even with a +20% power limit. There was really no point doing any of that with that card.

 

[rv8000]

220 watts and 304 watts seems like a very large range between an non-XT and XT card...

As far as we know the core clock is massively limited on the 9070, pretty sure some packaging/spec leaks were putting a 500-600mhz gap between the two. 9070XT is probably being pushed hard at that power envelope.

 

[alwayssts]

I don't know. I'm not big on overclocking anyway. Playing with the clocks only made them a bit more unstable even with a +20% power limit. There was really no point doing any of that with that card.

Maybe your card was a dud.
7800xt was literally *the* card to overclock from the last several years...because it's clock was freaking 2430mhz on a process that does 2700mhz pretty much minimum, and xt likely binned to 2800mhz.
It's the same how 7900xt's literally clock the same as many Blackwell (~3200mhz). W1zard only shows it in *some* reviews, because he didn't always use the extra 15% power. At stock, only 2800 or so.
Ask anyone that owns an xtx what they run at. That's literally the point of what AMD does....I don't get why people don't understand this.
Even nVIDIA expects this chip to clock to 3460mhz...I won't explain it bc it's boring bc math but just watch.

 

[3x0]

Maybe your card was a dud. 7800xt was literally *the* card to overclock from the last several years.

Don't let the TPU and other reviews fool you, they're not testing for 24/7 stability. I can replicate their 3DMark "stability" tests, but at 2700MHz and 1100mV (a much more conservative OC/UV than reviewers) mine crashed after 30 hours of Battlefield 2042. I'm not the only one, several others have experienced crashing after absurd amounts of time in game.

Sorry for OT.

 

[alwayssts]

Don't let the TPU and other reviews fool you, they're not testing for 24/7 stability. I can replicate their 3DMark "stability" tests, but at 2700MHz and 1100mV (a much more conservative OC/UV than reviewers) mine crashed after 30 hours of Battlefield 2042. I'm not the only one, several others have experienced crashing after absurd amounts of time in game.

Sorry for OT.

I believe it. Extreme clocks are not always 100% stable for hours on end, but you get my point. It was stock clocked very low (to sell cheap and for product segmentation, not bc capability).

 

[AusWolf]

Maybe your card was a dud.
7800xt was literally *the* card to overclock from the last several years...because it's clock was freaking 2430mhz on a process that does 2700mhz pretty much minimum, and xt likely binned to 2800mhz.
It's the same how 7900xt's literally clock the same as many Blackwell (~3200mhz). W1zard only shows it in *some* reviews, because he didn't always use the extra 15% power. At stock, only 2800 or so.
Ask anyone that owns an xtx what they run at. That's literally the point of what AMD does....I don't get why people don't understand this.
Even nVIDIA expects this chip to clock to 3460mhz...I won't explain it bc it's boring bc math but just watch.

I've seen lots of people clocking their XTX-es just shy of 3 GHz. I haven't seen many 7800 XTs running high clocks, though. But yeah, my sample is a total of 1 card, so I don't know.

As far as we know the core clock is massively limited on the 9070, pretty sure some packaging/spec leaks were putting a 500-600mhz gap between the two. 9070XT is probably being pushed hard at that power envelope.

The XT could be way too high up on the efficiency curve. Maybe you can limit it to 280, or even 260 W without losing any perceptible performance, like with a lot of high-end cards these days.

 

[alwayssts]

I've seen lots of people clocking their XTX-es just shy of 3 GHz. I haven't seen many 7800 XTs running high clocks, though. But yeah, my sample is a total of 1 card, so I don't know.


The XT could be way too high up on the efficiency curve. Maybe you can limit it to 280, or even 260 W without losing any perceptible performance, like with a lot of high-end cards these days.

That's exactly what I'm saying, friend. It was power limited so it wouldn't clock to 3ghz. 7700xt, a literal SALVAGE DIE of 7800xt, will clock that high bc not limited as much artificially (wrt clocks).
Look at the chart I just linked, 7700xts are included. Why does a WORSE PRODUCT on the SAME DIE clock higher? Because the 7800xt is (generally) power limited. Pure and simple.
So this chip (and 7900xt which was much more expensive) look good.
I know I've lost the war to Project Greenlight (which is literally when all this shit started...nVIDIA got pissed at AIBs for allowing higher-voltage/clocks which hurt margins/future products) but i'm still bummed.
People don't remember when we could actually do what we wanted with our hardware without having to go through a bunch of hoops. Huang won the war of attrition, this much is clear.
I just don't want AMD to do the same.
They've always had *some* bios limit, and now software power limit, but they've gotten more and more constrained.
Both on GPUs (7800xt is an example, so was 7900GRE with no memory slider until people complained, and then they literally lied through their teeth to get away with it by claiming it was a 'bug').
And even CPUs (I don't want to get into that bc it's more complicated).
What they are trying to do is LITERALLY what you said you don't want. They want to segment things as much as possible and make sure there is no overlap in current/future products.
WE WANT THAT OVERLAP because that helps us survive an upgrade cycle. Sometimes they allow it simply so they can match higher-end nVIDIA products (like 7800xt vs AD104 w/o discouraging not buying a 7900xt.)

Marketing is kinda complicated for some people, I get it, but the thing is they're tightening the noose...and it's not only bad business, it's also *very* stupid given their competitor and their market share.

At some point, some people will realize why we can no longer have nice things; why the whole market will be like what nVIDIA does.
It's because younger people will let companies take them from us bc they don't understand how things work.
It has taken Huang a LONG time to try to ween people like me out from the scene, but I still ain't dead yet.

 

[Assimilator]

AMD Radeon RX 9070 XT & RX 9070 GPU Performance Leaks Out: Comparable To RX 7900 GRE & RX 7800 XT

The first performance benchmarks of AMD's RDNA 4-based Radeon RX 9070 XT & 9070 GPUs have leaked out at Geekbench.

wccftech.com

Geekbench oh boy

 

[Zazigalka]

tbp is the total of gpu+memory+all circutry, right ?

 

[TheGeekn°72]

Accurate.

I'm pretty f'in toasted on talking about it until we know everything about it. (For me at least) it's fun to speculate and try to understand what they may do with the product, but this has been excessive.

I mean, this is our hobby and it's what we (and I) do...but sheesh. This prolonged escapade has been tiring for everyone, I think (and yes I do admit to my part in contributing to that).

Welp, looks like you/we're in luck ! GB6 leaks coming in to save the day ! Honestly, getting those kinds of benches at those freqs, I'm actually excited to see what would they clock at with the "game"/boost clocks !

I love the idea of interchangeable fans in the Magnetic Air series, although my attention is on the PowerColor Reaper this time, which seems to be the only truly dual-slot 9070 XT.

right after sending in my reply, I suddenly had the image of the white GPU with those red rimmed black fans, it'd look absolutely dope asf

 

[AusWolf]

What they are trying to do is LITERALLY what you said you don't want. They want to segment things as much as possible and make sure there is no overlap in current/future products.
WE WANT THAT OVERLAP because that helps us survive an upgrade cycle. Sometimes they allow it simply so they can match higher-end nVIDIA products (like 7800xt vs AD104 w/o discouraging not buying a 7900xt.)

What do you mean? The whole Nvidia 4070 (Super and up), 4080, 5070 and 5080 product stack is nothing more than just one big overlap. The same card with different names. This is what I don't want.

At some point, some people will realize why we can no longer have nice things; why the whole market will be like what nVIDIA does.
It's because younger people will let companies take them from us bc they don't understand how things work.
It has taken Huang a LONG time to try to ween people like me out from the scene, but I still ain't dead yet.

Looking on the bright side, at least we won't need to upgrade so often.

 

[alwayssts]

Geekbench oh boy

Yeah, looks about right.

 

[JustBenching]

At the same power and fab node on a single monolithic die as the 4080S, the 9070XT better be the same performance.

It should be faster. After all the 4080S is a rebranded 4060ti. I expect the 9070xt to beat it by ~30% in raster and be on par in RT.

 

[alwayssts]

What do you mean? The whole Nvidia 4070 (Super and up), 4080, 5070 and 5080 product stack is nothing more than just one big overlap. The same card with different names. This is what I don't need.


Looking on the bright side, at least we won't need to upgrade so often.

EXACTLY (wrt nvIDIA's stack).

This is why nVIDIA's parts (typically) overclock like crap. Blackwell clocks higher bc reasons (ram limitations, clock disparity to next products; 3nm; is huge). They might sell it to you with the ram later.
Before Rubin. Just to get another damn sale. The clock limits (with no ram limitation) will equal the stock clock perf of the next product (on 3nm). There's always room for a refresh with nVIDIA.
They sell every inch of performance (and then obsolete something by doing something as simple as increasing DLSS performance penalty greater than it's OC perf, it's og stock perf games' aim). Get it?

You are literally taking the exact opposite from what I'm saying than you should.

That's the whole point...we DO have to upgrade more often when we shouldn't HAVE TO because the former products are equally as capable (just locked down/deficient in some fixable way).
Then they are obsoleted by software demands outside the requisites of games (read: DLSS) and software relegated to only newer products.
It's soo fucking absurd and obvious, I don't even know how to make it any more clear. Do people not get it, or do they just not care?

*sigh*

I still love ya, but forget it. I really am not trying to fight with you; I'm trying to help you understand what's happening.

It should be faster. After all the 4080S is a rebranded 4060ti. I expect the 9070xt to beat it by ~30% in raster and be on par in RT.

LOLWUT? They'll probably be pretty even, if even that. I don't know where you guys think raster performance comes from?

8192 (let's pretend it isn't limited to ~7508 bc 64 ROPs) @ 2970-3100 = 9728 @ 2500-2610mhz. Stock clock of 4080 is 2730mhz. 4080s 10240 @ ~2775.

Geekbench implies it's clocked slightly higher when boosting....around where I thought it might. ~3200mhz or so? I guess it might be the TMUs, cache, or some other improvement. These things aren't magic, tho.

 

[TheGeekn°72]

This is why nVIDIA's parts (typically) overclock like crap. Blackwell clocks higher bc reasons (ram limitations, clock disparity to next products; 3nm; is huge).

isn't Blackwell on TSMC 4N ? I know it's an Nvidia specific node but that's still 4nm, right ?
and while RTX50 really thirsts for extra RAM, GD7 should provide enough bandwidth, no ? (referring to the "ram limitation" comment)

 

[alwayssts]

isn't Blackwell on TSMC 4N ? I know it's an Nvidia specific node but that's still 4nm, right ?
and while RTX50 really thirsts for extra RAM, GD7 should provide enough bandwidth, no ? (referring to the "ram limitation" comment)

It's complicated to explain. nVIDIA tuned 5nm for smaller die/tighter voltage/lower-clocks (lower leakage) and called it '4N' bc...well, because marketing. Also, small die equals more money. It's 5nm.
Current Blackwell products are closer to the original clock aim of general 5nm but utilizing the density improvement of 4nm. Basically, they always chose density (because then it's cheaper to make!)

5070 (and probably N48) likely use the performance potential (not density) to a much greater extent, but higher leakage (worse power efficiency). If they want, N48 could use a *ton* of power. But need bw.
Like I've said, nVIDIA's site implies 5070 is ~3150mhz or so, die size implies it's aimed for 3500mhz. The aim of N4P is 3460-3700mhz....but GPUs generally clock a little lower bc lots of units/big chips.
They are (*probably*) going to try to sell you a 4070ti with a different config/clock (less units, higher clock). This way they can excuse it being cheaper without devaluing products w/ more cores.
Basically, they think people are stupid. I am not going to give my opinion on that bc Huang has billions of dollars to prove what they do works.

As you see with the current blackwell, these libraries have ranges. 3165mhz might be the high for the denser library, but the low/easier yield for the higher-performance/less dense library.
Current B20x clocks similar to Navi3 when it's not power constrained, but N4P has a better voltage curve (can clock higher regardless even if dense, but higher over-all usable voltage before power goes nuts)
AMD chooses performance almost every time (make the most out of a small/efficient design). nVIDIA chooses density/power (often more units which use less power than proportional higher clocks.)
The reason why AMD does this is because it's typically cheaper to get the same performance this way, but it uses more power. nVIDIA's products are very concise bc...I don't want to explain it. Too much writing.
The point is they know exactly what they're doing...wayyy in advance...and planned obsolescene is in fact provabley real. I've been trying to explain all this crap, but it's....it's a lot for people to digest, I think.
Huang/nVIDIA are very good marketing people, perhaps the best on the planet, and in-fact he himself is almost-certainly a genius. I, however, am apparently more in-tune to spotting this stuff than most.

Sorry if I didn't write that coherently and/or it comes across wrong. I'm tired.

The ram thing is sooo....arrrgghhh. nVIDIA is trying to sell it that way, yes. They are *trying* to do that. But ram bandwidth is not the same as having that buffer. It's a long conversation...

 

[Daven]

Welp, looks like you/we're in luck ! GB6 leaks coming in to save the day ! Honestly, getting those kinds of benches at those freqs, I'm actually excited to see what would they clock at with the "game"/boost clocks !

Old news and debunked:

Alleged AMD Radeon RX 9070 XT nearly matches RX 7900 XTX in leaked Furmark 4K benchmark - VideoCardz.com

RX 9070 XT leak: Furmark is not just for stress testing It looks like AMD is preparing a surprise. AMD Radeon RX 9070 series are set to launch in two weeks, and in just one week AMD will disclose full details on the architecture. Gamers should look forward to full specs, performance, and...

videocardz.com

 

[Juventas]

And it's been confirmed that the RTX 5070 is 250 W TDP? Assuming the RX 9070 is performance competitive, this would make it more efficient as well?

 

[Zazigalka]

They sell every inch of performance (and then obsolete something by doing something as simple as increasing DLSS performance penalty greater than it's OC perf, it's og stock perf games' aim). Get it?

to be precise, the performance penalty of dlss4 on ampere/turing comes from the RT/PT denoiser (ray reconstruction), not the dlss upscaler.